Holt-Winters exponential smoothing model with 52-week seasonality.
Select training start and cut-off dates to see the forecast vs actuals.
About this forecast
The chart shows the CDC's national
Wastewater Viral Activity Level
(WVAL) for SARS-CoV-2 and Influenza A, one value per week since January 2022. A bundled historical
snapshot is merged with the live
CDC data feed,
which takes precedence where the two overlap.
Model
- Values are Box–Cox transformed, with the exponent chosen by maximum likelihood, to stabilize the variance.
- An additive Holt–Winters model with a damped trend and 52-week seasonality is fit to the transformed series.
Smoothing parameters and initial states are chosen by a bounded Powell-style search that minimizes squared one-step errors.
- The forecast runs 52 weeks past the cut-off date and is transformed back to the activity-level scale.
- Everything runs in your browser: standard-library Python on Pyodide
in a Web Worker, with no server and no third-party Python packages.
Backtesting with the sliders
- The gray slider sets the first week of training data and the blue slider sets the last (the cut-off).
- Training always spans at least 104 weeks, the two seasonal cycles needed to initialize the model.
When the window is at that minimum, dragging either slider pushes the other along.
- History before the start date is drawn in light gray; actual values after the cut-off are dotted, for comparison with the forecast.
Confidence band
The shaded band spans the 25th to 75th percentiles of up to 500 simulated paths, each adding Gaussian errors on the
transformed scale; simulation stops early once those percentiles change by less than 5% on average.
For COVID-19, the upper edge is capped at an activity level of 30 for display.
Limitations
- The band reflects only week-to-week noise. It ignores uncertainty in the fitted parameters and in the size of the
seasonal peak, so it is usually too narrow, especially in the weeks before a winter peak.
- The January 2022 Omicron surge is far larger than any later winter peak. When it is in the training window, it
inflates the model's January seasonal effect; try a training start date after spring 2022 to compare.
- COVID-19 waves do not follow a strict 52-week calendar, so the timing of forecast peaks is approximate.
- This is an exploratory statistical model, not a public health forecast. For official information, see the
CDC National Wastewater Surveillance System.